Organizations: Zhejiang University · Shanghai AI Laboratory · Shanghai Jiao Tong University · Tsinghua University · The University of Hong Kong · The Chinese University of Hong Kong
Recent monocular 3D Gaussian Splatting (3DGS) streaming reconstruction methods have achieved impressive performance by balancing reconstruction quality and efficiency. However, extending these frameworks to underwater scenes remains challenging due to severe visual degradation, such as light attenuation and scattering, which degrades camera pose tracking and distorts scene geometry. To address these challenges, we propose AquaFlow, a monocular Gaussian Splatting streaming reconstruction framework for efficient and high-fidelity underwater reconstruction. Specifically, AquaFlow fine-tunes a 3D vision foundation model on large-scale underwater data for robust pose and pointmap estimation, and introduces a medium-guided incremental Gaussian initialization strategy for streaming mapping. Furthermore, we develop a streaming-compatible hybrid scene representation that integrates structured, distance-conditioned neural Gaussians with a physics-inspired optical model to compensate for underwater image formation effects, enabling accurate scene reconstruction. We evaluate AquaFlow on a comprehensive dataset of 62 diverse underwater trajectories, collected from both public benchmarks and in-the-wild web videos across various scales. Extensive experiments demonstrate that AquaFlow achieves state-of-the-art tracking and rendering performance, reducing average localization error by 13.2% and improving PSNR by 4.74 dB compared to WaterSplat-SLAM.
Figures & tables
Fig. 1: Overview of the AquaFlow framework. Given a long, continuous underwater video sequence, AquaFlow performs online streaming 3D reconstruction, rapidly estimating camera poses, dense pointmaps, a highly compact 3D Gaussian map, and physical medium parameters. To achieve this, AquaFlow seamlessly integrates domain-adapted underwater 3D geometric priors, a medium-guided Gaussian initialization scheme, and a hybrid scene representation combining structured neural Gaussians with physics-based medium modeling, ultimately enabling accurate subsea structural recovery and photorealistic novel view synthesis.
Fig. 2: Overview of the AquaFlow streaming 3D reconstruction pipeline. Given incoming monocular underwater frames, AquaFlow first predicts dense pointmaps and descriptors using a domain-adapted 3D foundation model to establish reliable pixel correspondences for relative pose optimization. Based on match counts and parallax, frames are dynamically categorized into keyframes, mapping frames, or common frames. Keyframes and mapping frames trigger medium-guided incremental initialization , where an online learned physical medium model removes attenuation and backscatter to yield clean base colors and structure-aware Gaussian seeds. Finally, these primitives are organized into a distance-aware structured neural Gaussian representation and decoded via the LGPD module, which is jointly optimized alongside explicit physical medium modeling for high-fidelity rendering.
Fig. 3: Overview and taxonomy of our proposed underwater evaluation benchmark. The dataset spans broad physical spatial scales and rich topological diversity, encompassing characteristic benthic environments such as shallow sandy beds, canyons, coral reefs, rocky terrains, submerged caves, and artificial ruins. Furthermore, it incorporates diverse camera trajectories (e.g., straight paths, loops, and erratic 3D motions) across a wide range of frame lengths: 14 short sequences ( 0–50 frames), 13 medium-short sequences ( 50–100 frames), 20 medium sequences ( 100–200 frames), 3 long sequences ( 200–400 frames), and 12 extended-range sequences ( >400 frames) to comprehensively assess online tracking and long-term mapping stability.
Dataset
Pairs
Type
Scene Content
TartanAir-Ocean
76,387
Syn.
Seafloor, terrain, industrial, wrecks
MIMIR-UW
44,721
Syn.
Seafloor, algae, industrial, ruins
UWStereo
26,611
Syn.
Corals, vessels, structures, ROVs
FLSea-stereo
6,576
Real
Seafloor, canyons, corals, artifacts
sweet-corals
69,978
Real
Corals, complex biomes & structures
Total
224,273
Mixed
Varied benthic biomes and subsea structures
TABLE I: Composition of underwater training image pairs. The datasets cover diverse synthetic and real underwater scenes.
Method
Canyons
RedSea
UW-Stereo-VI
SeaThru-NeRF
S-UW
Internet Data
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
UW-GS
27.33
0.916
0.288
23.78
0.809
0.316
27.87
0.903
0.366
25.79
0.858
0.260
20.55
0.733
0.279
23.56
0.786
0.311
OnTheFly-NVS
27.20
0.894
0.317
22.37
0.699
0.387
29.32
0.888
0.385
22.12
0.719
0.324
19.51
0.623
0.335
23.42
0.746
0.339
HI-SLAM2
29.88
0.898
0.269
23.66
0.716
0.448
27.69
0.856
0.423
25.33
0.745
0.471
21.42
0.685
0.367
25.13
0.769
0.369
MonoGS
27.83
0.875
0.460
20.89
0.644
0.708
25.87
0.861
0.480
13.82
0.524
0.589
9.70
0.261
0.667
16.47
0.607
0.623
S3PO-GS
29.62
0.910
0.369
22.47
0.669
0.562
29.96
0.900
0.450
19.48
0.609
0.493
19.48
0.569
0.459
22.82
0.743
0.491
TABLE II: Quantitative reconstruction results on real-world underwater datasets. We report PSNR, SSIM, and LPIPS. Higher PSNR/SSIM and lower LPIPS indicate better performance. The best and second-best results are in bold and underline .
Fig. 4: Qualitative comparisons of AquaFlow against baselines [ 6 , 46 , 22 , 17 , 33 ] . The top three rows highlight AquaFlow’s superior modeling capability for distant underwater scenes and intricate seafloor details (e.g., ground markings and discarded vehicle textures). The bottom three rows demonstrate our strength in recovering complex subsea structural topologies (e.g., underwater caves, shipwreck ruins and canyons), where domain-adapted 3D geometric priors and appearance representations effectively preserve sharp object boundaries and detailed structures compared to baseline methods.
Method
FLsea_canyons
FLsea_redsea
underwater-stereo-vi
seathruNerf
S-UW
Internet data
Time ↓
FPS ↑
#GS(k) ↓
Time ↓
FPS ↑
#GS(k) ↓
Time ↓
FPS ↑
#GS(k) ↓
Time ↓
FPS ↑
#GS(k) ↓
Time ↓
FPS ↑
#GS(k) ↓
Time ↓
FPS ↑
#GS(k) ↓
UW-GS
1896.0
0.26
952.0
4174.0
0.12
3120.0
557.6
0.31
750.9
227.5
0.10
707.8
245.2
0.10
1032.0
557.0
0.06
1132.9
ARTDECO
498.4
1.00
464.2
593.3
0.84
1006.4
212.8
0.74
733.9
50.6
0.43
917.4
39.8
0.60
1214.2
115.0
0.92
705.2
WaterSplat-SLAM
1562.9
0.32
13.8
1029.0
0.48
26.9
757.8
0.21
7.4
75.9
0.29
33.8
88.5
0.27
37.1
147.6
0.71
750.0
Ours
467.1
1.08
610.5
709.4
0.69
975.7
207.2
0.76
581.9
38.5
0.56
790.2
47.3
0.51
1157.8
153.9
0.68
758.6
TABLE III: Quantitative efficiency and Gaussian number comparison. We report scene training time ( Time (s)↓ ), training throughput ( FPS↑ ), and Gaussian count ( #GS (k)↓ ). The best and second-best results are in bold and underline , respectively.
Method
Canyons
RedSea
UW-Stereo-VI
SeaThru-NeRF
S-UW
Internet
UFEN-SLAM
1.107
1.226
0.978
1.637
1.570
1.477
DROID-SLAM
0.935
0.737
2.969
0.376
0.069
1.057
MASt3r-SLAM
0.978
1.357
0.872
1.579
1.270
0.997
VGGT-SLAM
3.079
2.747
1.965
0.237
0.469
2.261
VGGT-SLAM2
0.732
1.206
0.520
0.240
0.470
1.339
ARTDECO
0.559
0.991
0.325
0.372
0.245
0.593
TABLE IV: Quantitative localization (trajectory) results on underwater datasets. We report ATE RMSE (m) ↓ . The best and second-best results are in bold and underline , respectively.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
GS(k)
Full model
34.47
0.953
0.209
601.6
w/o UW finetuning
33.41
0.943
0.247
608.3
w/o medium init.
33.91
0.949
0.223
406.8
w/o dist. embedding
32.56
0.943
0.232
629.6
w/o LGPD
33.92
0.951
0.214
653.2
TABLE V: Ablation study on the FLSea-Canyons dataset. We report PSNR, SSIM, and LPIPS. Higher PSNR/SSIM and lower LPIPS indicate better reconstruction quality.
Fig. 5: Qualitative ablation comparison on rendered color and depth maps. Compared to baseline variants lacking specific modules (e.g., medium initialization, underwater fine-tuning, distance encoding, or LGPD), disabling any component introduces structural geometric errors or scene blurring. In contrast, our full model faithfully recovers both accurate 3D geometry and realistic underwater appearance.
Department of Computer Science & Engineering and the Minnesota Robotics Institute (MnRI), University of Minnesota–Twin Cities, Minneapolis, MN 55455 USA